/lev-audit [daily|weekly]

Grep is only allowed to be the last word where the rulebook says keep-mechanical. Everything else is a rule an agent reads, checks with evidence, and signs.

Inputs

• ~/lev/dna/gates/guard-rulebook.yaml — families, guards, dispositions, agent_evidence per family.
• ~/lev/.lev/validation-gates.yaml, ~/lev/dna/gates.yaml — gate ids the rulebook points at.
• Scripts live in ~/lev/tooling/scripts/.

daily (default)

1. For every guard with disposition: keep-mechanical, run its script from ~/lev. Also run pnpm axes-check. Guards may carry an args field; pass it (shell-globbed, from ~/lev) to the script.
2. Append one row per guard to ~/lev/.lev/state/audit/<YYYY-MM-DD>-daily.jsonl ({family, guard, disposition, mechanical, evidence}).
3. Any failure: open a bd issue audit: <family>/<guard> failed with the command output. Do not fix.

weekly

1. Do the daily run first.
2. For every guard with mechanical+agent-review or agent-rule-only, delegate one worker per family (sonnet or /codex-runner, never opus). Give the worker the family's rule and agent_evidence text verbatim. The worker must: run the script (if any) as a pre-filter, then gather the evidence the family names (import graph, live probe, doc diff, sampled plugins), and return pass|fail per guard with file:line or command output as evidence. A pass without evidence is a fail.
3. Append rows to <date>-weekly.jsonl with agent and evidence filled.
4. Every agent: fail → bd issue audit: <family>/<guard>: <one-line finding>.
5. Reply with a table: family · guards · mechanical · agent · issues opened. No prose beyond one line per failure.

Rules

• Never edit code during an audit. Findings become bd issues.
• Never trust a regex pass for a judgment rule; say n/a for mechanical when the family says the script proves nothing.
• Unknown guard mechanism → mark tbd and add a row asking for classification; do not guess.
• Schedule: add to ~/lev/.lev/flows/lev-mgmt.schedule.yaml (daily 09:00 lev-audit daily, Monday lev-audit weekly) — the same tick parity-scorecard uses.